avfilter/asubprocess: optimize 32 bit i/o on little endian systems

Signed-off-by: Stefan Westerfeld <stefan@space.twc.de>

libavfilter/af_asubprocess.c

@@ -310,42 +310,95 @@ static int config_input(AVFilterLink *inlink)
     return 0;
 }
 
-static int filter_frame(AVFilterLink *inlink, AVFrame *in)
+static void read_samples(ASubProcessContext *s, int32_t *data, int count)
 {
-    AVFilterContext *ctx = inlink->dst;
-    ASubProcessContext *s = ctx->priv;
-    AVFilterLink *outlink = ctx->outputs[0];
-    AVFrame *out;
-    int ret = 0, nb_samples;
+    int sample_size = s->in_bit_depth / 8;
+    assert(in->nb_samples <= NB_BUFFER_SAMPLES);
+#if !HAVE_BIGENDIAN
+    if (s->out_bit_depth == 32)
+    {
+        /* optimized case: on little endian systems we don't need any conversion */
+        sp_read(s->sp, (char *)data, count * sample_size);
+        return;
+    }
+#endif
+    sp_read(s->sp, s->sample_buffer, count * sample_size);
+    if (s->out_bit_depth == 8)
+    {
+        int8_t *in = s->sample_buffer;
+        for (int k = 0; k < count; k++)
+            data[k] = (in[k] << 24) + 0x80000000;
+    }
+    if (s->out_bit_depth == 16)
+    {
+        int16_t *in = s->sample_buffer;
+        for (int k = 0; k < count; k++)
+            data[k] = AV_RL16(in + k) << 16;
+    }
+    if (s->out_bit_depth == 24)
+    {
+        uint8_t *in = s->sample_buffer;
+        for (int k = 0; k < count; k++)
+        {
+            data[k] = AV_RL24(in) << 8;
+            in += 3;
+        }
+    }
+    if (s->out_bit_depth == 32)
+    {
+        int32_t *in = s->sample_buffer;
+        for (int k = 0; k < count; k++)
+            data[k] = AV_RL32(in + k);
+    }
+}
 
-    int k = 0;
-    bool gen_output;
-    int32_t *xd = (int32_t *)in->data[0];
+static void write_samples(ASubProcessContext *s, int32_t *data, int count)
+{
     int sample_size = s->in_bit_depth / 8;
-    int todo = in->nb_samples * inlink->ch_layout.nb_channels;
-    assert (in->nb_samples <= N_SAMPLES);
+    assert(in->nb_samples <= NB_BUFFER_SAMPLES);
+
+#if !HAVE_BIGENDIAN
+    if (s->out_bit_depth == 32)
+    {
+        /* optimized case: on little endian systems we don't need any conversion */
+        sp_write(s->sp, (char *)data, sample_size * count);
+        return;
+    }
+#endif
     if (s->in_bit_depth == 16)
     {
         int16_t *out = s->sample_buffer;
-        for (k = 0; k < todo; k++)
-            AV_WL16 (out + k, xd[k] >> 16);
+        for (int k = 0; k < count; k++)
+            AV_WL16(out + k, data[k] >> 16);
     }
     if (s->in_bit_depth == 24)
     {
         uint8_t *out = s->sample_buffer;
-        for (int k = 0; k < todo; k++)
+        for (int k = 0; k < count; k++)
         {
-            AV_WL24 (out, xd[k] >> 8);
+            AV_WL24(out, data[k] >> 8);
             out += 3;
         }
     }
     if (s->in_bit_depth == 32)
     {
         int32_t *out = s->sample_buffer;
-        for (k = 0; k < todo; k++)
-            AV_WL32 (out + k, xd[k]);
+        for (int k = 0; k < count; k++)
+            AV_WL32(out + k, data[k]);
     }
-    sp_write (s->sp, s->sample_buffer, sample_size * todo);
+    sp_write(s->sp, s->sample_buffer, sample_size * count);
+}
+
+static int filter_frame(AVFilterLink *inlink, AVFrame *in)
+{
+    AVFilterContext *ctx = inlink->dst;
+    ASubProcessContext *s = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
+    AVFrame *out;
+    int ret = 0, nb_samples;
+    bool gen_output;
+
+    write_samples(s, (int32_t *)in->data[0], in->nb_samples * inlink->ch_layout.nb_channels);
     if (s->eof)
     {
         while (!sp_write (s->sp, 0, 0))
@@ -398,43 +451,12 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
             avail = NB_BUFFER_SAMPLES;
         if (avail > 0)
         {
-            int32_t *xd;
-            int todo;
             out = ff_get_audio_buffer(outlink, avail);
             if (!out) {
                 av_frame_free(&in);
                 return AVERROR(ENOMEM);
             }
-            sp_read (s->sp, s->sample_buffer, avail * sample_size * inlink->ch_layout.nb_channels);
-            todo = avail * inlink->ch_layout.nb_channels;
-            xd = (int32_t *)out->data[0];
-            if (s->out_bit_depth == 8)
-            {
-                int8_t *in = s->sample_buffer;
-                for (int k = 0; k < todo; k++)
-                    xd[k] = (in[k] << 24) + 0x80000000;
-            }
-            if (s->out_bit_depth == 16)
-            {
-                int16_t *in = s->sample_buffer;
-                for (int k = 0; k < todo; k++)
-                    xd[k] = AV_RL16 (in + k) << 16;
-            }
-            if (s->out_bit_depth == 24)
-            {
-                uint8_t *in = s->sample_buffer;
-                for (int k = 0; k < todo; k++)
-                {
-                    xd[k] = AV_RL24 (in) << 8;
-                    in += 3;
-                }
-            }
-            if (s->out_bit_depth == 32)
-            {
-                int32_t *in = s->sample_buffer;
-                for (int k = 0; k < todo; k++)
-                    xd[k] = AV_RL32 (in + k);
-            }
+            read_samples(s, (int32_t *)out->data[0], avail * inlink->ch_layout.nb_channels);
             ret = ff_filter_frame(outlink, out);
         }
         else {